One source. Native agent files. A receipt that does not pretend they loaded.
Edit the project rule, click once, and inspect the generated CLAUDE.md, AGENTS.md, or .cursorrules. Then change the generated copy and make the write-back policy explicit. Everything runs locally in this browser; nothing is uploaded or written.
Canonical source
Change “Run tests before merging” below, then generate again.
Native output
Target-edit triage
Real workflows do not stay one-way: a human or agent may edit a generated target. Change the observed file below, choose a policy, and inspect the decision. “Promote” means review the diff for the canonical source; it never silently makes the target authoritative.
Configured is not loaded — and loaded is not useful
Same guidance, same tiny task, same runtime version. Lane A uses a filename the target silently ignores; lane B uses its native filename and observes the source hash at the request boundary. Token count describes cost, not usefulness, so each lane keeps verification and accepted outcome separate.
A · configured, ignored
B · native filename, observed
Measurement motivation: Systima's API-boundary study observed Claude Code 2.1.207 ignoring AGENTS.md but loading the same content as CLAUDE.md. Projection tools such as agentctl can prove delivery to a native-session boundary; destination observation still needs its own evidence.
Boundary receipt
This proves what this preview transformed. It does not claim Claude, Cursor, or another runtime loaded the file.
Try the real CLI
npx --yes --package github:caioribeiroclw-pixel/pluribus#v0.3.52 pluribus init --dry-run --name "Acme API" --description "A TypeScript billing service" --tools claude,cursor,openclaw
The pinned GitHub release avoids the current npm release lag. --dry-run previews without writing files.